OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling
Graded visual feedback outperformed binary checks in text-to-image tests; pairwise verification proved more accurate when ground-truth images were available.
OmniVerifier-C1 breaks complex prompts or environmental rules into fine-grained checks and scores each component. Its Decoupled RL training combines binary judgments on individual sub-questions with an overall alignment score. The authors report that these structured outputs improve generalization to out-of-domain multimodal critique tasks. They also provide theoretical support for pairwise verification and validate the approach on structured world-modeling tasks including Maze and Sudoku.
HF Daily Papers' note
OmniVerifier-C1 breaks complex prompts or environmental rules into fine-grained checks and scores each component. Its Decoupled RL training combines binary judgments on individual sub-questions with an overall alignment score. The authors report that these structured outputs improve generalization to out-of-domain multimodal critique tasks. They also provide theoretical support for pairwise verification and validate the approach on structured world-modeling tasks including Maze and Sudoku.
HF Daily Papers' note
score 5